Performance analysis of an integrated eye gaze tracking / electromyogram cursor control system
Craig A. Chin, Armando Barreto, J. Gualberto Cremades, Malek Adjouadi · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '07) · doi:10.1145/1296843.1296888
Summary
Chin, Barreto, Cremades, and Adjouadi's Assets '07 demonstration paper from Florida International University evaluates a hybrid cursor-control system that combines eye-gaze tracking (EGT) with electromyogram (EMG) input from facial muscles, targeted at users with motor disabilities who cannot operate a conventional mouse. The authors motivate the hybrid by the classic complementary weaknesses of each modality: EGT is fast but prone to unsteadiness and 'Midas touch' (unintended selection whenever the user looks at a target) because the eye never really stops moving, while EMG-driven step-by-step cursor control is precise and gives an unambiguous 'click' (a jaw clench), but slow. The EGT subsystem uses an ASL Model RHS P/T6 eye tracker sampling point-of-gaze at 120 Hz; a fixation is registered when the standard deviation of x,y gaze coordinates across a 100-ms window falls below the threshold for 0.5 degrees of visual angle, and an EGT-only click is issued after a 350-ms dwell. The EMG subsystem uses surface electrodes on the right and left temples (temporalis), forehead (frontalis), and brow (procerus), and distinguishes which muscle fired by analysing the Mean Power Frequency of each channel's power spectral density — needed because muscle signals bleed across electrodes due to volume conduction in the head. Direction commands move the cursor in discrete ±Δx/±Δy steps; a simultaneous jaw clench issues a click. The integrated system uses the EGT for gross cursor positioning (teleporting to a 'qualified fixation') and the EMG for fine-grained correction and clicking. Two point-and-click experiments compared EGT-alone, the EGT/EMG hybrid, and a standard mouse.
Key findings
In Experiment 1 (30 able-bodied participants, 10 per condition, 72 Fitts-style point-and-click trials each), mean completion time was 0.94 s for the mouse, 3.07 s for EGT alone, and 4.68 s for EGT/EMG — the hybrid was significantly slower than both alternatives. However, mean error rate (off-target clicks per trial) was 0.01 for the mouse, 0.13 for EGT/EMG, and 3.98 for EGT alone — a roughly thirty-fold reduction in selection errors when EMG was added to EGT, bringing error rate essentially to mouse-level (difference not statistically significant, p = 0.206). Experiment 2 (15 participants, 'Y'/'N' go/no-go targets) confirmed the accuracy story: mean error rate per trial was 0.396 for EGT versus 0.017 for EGT/EMG, a highly significant drop (Wilcoxon signed-rank). The authors frame the trade-off honestly: EMG-augmented cursor control buys precision and click reliability at the cost of speed, and the headline finding is that the hybrid effectively eliminates the Midas Touch problem while keeping EGT's ability to jump quickly to the vicinity of a target. All participants were able-bodied; testing with people with motor disabilities is listed as future work.
Relevance
For accessibility practitioners designing mouse-alternatives for users with motor disabilities — ALS, high-level spinal cord injury, severe cerebral palsy, locked-in syndrome — this paper is a useful data point on the speed-versus-accuracy trade-off between pure gaze and multimodal gaze input. The Midas Touch problem remains unsolved in gaze-only systems and is the single most-cited reason blind-deployment of gaze control fails in practice; the idea of using a separate, slower, but unambiguous confirmation channel (jaw clench, sip-and-puff, brow raise, single switch) is a design pattern that continues to matter for modern eye-gaze AAC products, VR gaze interaction, and emerging brain-computer interfaces. The authors' treatment of volume conduction — the observation that surface EMG channels are not cleanly separated and require spectral discrimination to attribute a contraction to a specific muscle — is a non-obvious implementation lesson for anyone building facial-EMG input. Limitations are real: no evaluation with actual motor-impaired participants, no long-term learning study (the authors speculate that extended use would narrow the speed gap but do not test it), and the 2007 hardware is obsolete. The design principles remain directly applicable to modern multimodal-gaze products such as Tobii Dynavox, Control Bionics NeuroNode, and emerging AI-assisted gaze systems.
Tags: eye tracking · electromyogram · cursor control · motor accessibility · multimodal input · mouse alternative · midas touch · alternative input · biosignals · Fitts law